Applications of Optimization and Machine Learning to Healthcare

Abstract

As the field of healthcare becomes increasingly data-driven, optimization and machine learning methods provide the scientific community and practitioners with powerful tools to extract insights from the data with potential to improve clinical practice and inform public health efforts. However, medical applications pose a set of unique challenges. Medical data is often limited since accumulating large amounts of patient data with labels is difficult, especially for rare conditions or hospital-specific tasks. Additionally, the high stakes healthcare environment requires development of reliable explainability methods to ensure the safe application of machine learning models. On the other hand, abundant unlabelled data on the population level provides unique opportunities in public health. In our work we leverage these opportunities and develop methods to address the above challenges. We first consider applications in radiation oncology and public health. We propose a novel optimization framework for multi-modality radiation therapy and discuss our work on leveraging unlabelled cell phone mobility data and manifold learning to gain insights into population behavior during the COVID-19 pandemic. We then proceed to the methodological part of our work. We propose a novel parameter-saliency explainability method for deep neural networks which can be used to analyze model mistakes. After that, we present our work on transfer learning with deep tabular models where in a realistic medical setting we show that representation learning with neural networks provides a definitive advantage over the traditionally dominant gradient boosted decision tree tabular methods when downstream data is limited. Finally, we present an application of deep tabular models to improve patient-specific quality assurance in radiation oncology.